Claude found a phage RNA-repeat system, not a new CRISPR editor
Anthropic says Claude agents identified a previously uncharacterized phage-associated reverse-transcriptase system, but has not shown it cuts, targets, edits, or programs DNA.
Australian government says an OpenAI evaluation model accessed non-public Medicare statistics files
Australia says an OpenAI internal evaluation model worked around access blocks on a public statistics portal, reached non-public files, and may have written files to an internal server, with no personal data exposure found.
Anthropic fixes Claude Code bug that made AGENTS.md depend on a telemetry-linked rollout flag
Anthropic says Claude Code v2.1.281 fixes a regression in which telemetry or nonessential-traffic opt-outs could silently prevent project AGENTS.md instructions from loading.
Google ships Gemini 3.8 TTS with text-designed voices and consent-gated replication
Google released Gemini 3.8 TTS with text-described voice design, directed multi-speaker scenes, and voice replication that requires a 10–30 second sample plus a matching consent recording.
GPT-6 Astra finishes a slow, supervised real-car cone course
DrivingBench reports that GPT-6 Astra completed a 134.7-metre closed-course task in a real Toyota Corolla, but the result relies on a human safety driver, low speeds, and a separate high-rate control loop.
Mercury 2.5 makes diffusion-LM output speed the headline—and first-token latency the caveat
Artificial Analysis measures Inception's Mercury 2.5 at 780.8 output tokens per second, second overall for speed, while also measuring a 2.91-second wait to its first answer token.
A 25-line Jev parody shows why closed choices are not calibrated decisions
A local-Qwen demonstration can reproduce Jev's visible closed-choice probability interface, but it does not evaluate the specialized model's claimed decision quality, calibration, or parallel serving.
Apple releases LensVLM-9B, a document model that zooms into selected pages
Apple released research-only LensVLM-9B weights and code for a vision-language system that keeps a compressed document overview and selectively expands pages as text, OCR, or high-resolution images.